Our study shows that although nursing home closures in the United States are a relatively rare event (approximately 5% of facilities over the past decade), they can be predicted with reasonable accuracy using machine learning on longitudinal data. We found that advanced sequence-based models perform better than traditional prediction approaches. In particular, recurrent neural network models (LSTM and bidirectional LSTM) achieved significantly higher discriminatory ability than baseline models such as logistic regression (LR) and random forest (RF). For example, our best LSTM model had an AUPRC of about 0.50 and a recall of 0.77. In comparison, the best-performing traditional model had ~0.314 and 0.34, respectively. This improvement suggests that incorporating temporal trends in facility performance provides important predictive signals in key metrics that are particularly more relevant to studies of a high-disequilibrium nature: AUPRC, precision, recall, and F-1 score.
This study builds on previous cross-sectional studies that identified factors associated with closure and shows that these factors can also prospectively identify high-risk facilities.4,10. We looked at the geographic aspects of closure risk. Rural facilities had a higher probability of closure, which is consistent with reports of “nursing home deserts” in rural counties where the closure of a single facility can result in loss of local access to care.4,5. Our results are consistent with research that closures tend to be concentrated in socio-economically disadvantaged areas. Fenn et al.27 Nursing home closures found to be more common in counties with higher poverty rates and larger minority populations, raising concerns about equity27. From a health equity perspective, this pattern means that communities already facing limited resources suffer disproportionately from the loss of facilities, further exacerbating disparities in access to care.
Additionally, our analysis identified several predictors of closure risk. Facilities that closed were more likely to have lower occupancy rates, smaller average censuses, rural locations, and signs of quality and care challenges (e.g., significantly higher rates of resident catheter use and lower overall star ratings) than those that remained open. These characteristics emerged as consistent risk factors in our model. Even facilities incorrectly flagged as high risk (false positives) by the model tended to share many of these risk characteristics with actual closed facilities, confirming that the model was capturing the true profile of vulnerability. Taken together, these findings provide new evidence that nursing home closures are not random or unpredictable, but rather predictable based on observable facility characteristics and trends, and that deep learning models leveraging longitudinal data provide superior performance in accurately identifying at-risk nursing homes.
Another important aspect of our approach is the emphasis on recall (sensitivity) when predicting this rare event. Given the high degree of imbalance (only ~5% of records represent closures), our modeling prioritized capturing as many true closures as possible, even at the expense of some false positives. As a practical matter, missing an impending closure (a false negative) is far more harmful to nursing home regulators and residents than flagging a facility that ends up remaining open (just as a false negative can be worse than a false positive in disease screening). By maximizing model recall, we aimed to be able to identify most of the facilities that would be closed in advance. In fact, our best model detected the majority of real closures, confirming that a recall-oriented strategy is suitable for this situation. The trade-off is that many false-positive predictions occur, but our analysis reveals that these false-positive facilities are typically “near misses” in that they exhibit many of the same risk factors as true closures. For example, a subset of facilities that the model incorrectly identified as likely to close were actually SFF candidate facilities that were on the SFF list or had well-documented quality and compliance issues. Although we found that 14 of the model's false-positive facilities were designated as SFF/SFFC facilities, and only 5 of them ultimately closed during the study, the inclusion of these facilities highlights that the model effectively captures markers of severe quality of care decline. Essentially, the “false alarms” in this model are often facilities exposed to significant hardship (e.g., low census scores, poor quality ratings, regulatory troubles) that have previously avoided closure, often due to temporary intervention or lack of nearby competitors. This observation indicates that not all poorly performing nursing homes will close. While some nursing homes manage to stay open, perhaps due to outside support or critical needs in the community, we find that those that close most often come from a cluster of troubled facilities. It also suggests that many false-positive facilities may remain at risk of closure in the long term unless the underlying issues are resolved. Overall, the consistency between our model's risk flags and known indicators of nursing home instability gives us confidence that the predictive patterns our model has learned are meaningful and not spurious. Additionally, the SFF program flags a limited number of facilities as underperforming, primarily based on quality and safety issues. The proposed model provides a more comprehensive view of nursing home performance and nursing home closure risk by employing a broader range of characteristics, including market, financial, resident, quality, and other facility characteristics.
These findings have several implications for state and federal regulators tasked with overseeing nursing homes. First, predictive analytics could be integrated into regulatory monitoring systems as an early warning tool for facility closure risk. Rather than waiting for official closure announcements or clear signs of financial failure, agencies such as CMS and state regulators could use model-based alert systems to proactively identify nursing homes that are showing clear signs of distress (e.g., persistently low occupancy rates, declining quality, etc.). Receiving a data-driven risk warning, perhaps a year in advance of a potential closure, gives regulators lead time to intervene or at least prepare. Second, policymakers can leverage predictions of closure risk to more efficiently target interventions and resources. Facilities reported to be at high risk may be offered support such as temporary financial relief, administrative support, and professional oversight to address deficiencies. These are measures that may stabilize the facility and prevent avoidable closures. When closure appears unavoidable, early identification allows residents to make a well-prepared transition plan (e.g., find an alternative location, notify family members) and minimize the disruption and negative health effects associated with sudden evacuation. Third, our model highlights geographic areas of concern (rural areas and areas with high uninsured rates), which can inform broader policy strategies on health equity and access. Regulators could prioritize maintaining services in these vulnerable areas, for example by encouraging new operators to enter “nursing home desert” counties or expanding home- and community-based services where nursing homes have closed. In this way, predictive modeling can support more equitable resource allocation and ensure that underserved communities are not left behind by unmonitored market forces.
This study has several limitations that are worth considering. First, there are limitations to the available data. Important facility-level variables, particularly from CMS cost reports (e.g., detailed financial performance indicators), contained significant missing values. These data gaps required imputation and may have reduced the model's accuracy in assessing financial hardship, an important precursor to closure. Improving the completeness and quality of such data in the future will enhance its predictive power. Second, nursing home closures are inherently an imbalanced classification problem, with only a small number of even low-performing facilities closing, which complicates model training. Although we applied SMOTE oversampling to address this imbalance, the underlying biased distribution still means that even if recall is high, positive predictive value is limited (i.e., many false positives). The gap between AUROC and more appropriate metrics such as AUPRC, precision, and recall, especially outside the closure sample set, indicates that the tradeoff between precision and recall still persists. To address this as much as possible, model training was directed towards optimizing recall performance. In practice, this means that model alerts must be interpreted carefully. Flagged facilities are not guaranteed to be closed, but rather should be considered high risk and in need of caution. Third, the best-performing models in our analysis were deep learning models (particularly BI-LSTM) that behaved as “black boxes.” These recurrent neural network models do not readily provide interpretable explanations as to why a particular facility is predicted to be at risk. This lack of transparency can pose challenges for adoption in policy settings where stakeholders may demand a clear basis for decision-making. Although we attempted to glean insight by comparing the properties of true closures and false-positive predictions, this post-hoc analysis is no substitute for built-in interpretability. Future research may consider using knowledge distillation models to balance predictive power and explainability, allowing regulators to act with confidence in model outputs.28,29. Finally, our analysis is based on data from 2011 to 2021. The long-term care sector is in flux, especially as closures have increased since 2020 due to the COVID-19 pandemic.2. Our model does not yet capture pandemic-era policy changes or the unprecedented burdens that have emerged in recent years (e.g., labor shortages, epidemic response costs). Therefore, the model may need to be retrained and recalibrated with data from 2021 and beyond to remain effective. Despite these limitations, we believe that our approach provides a strong proof of concept for predicting closures, and that the general patterns identified (e.g., occupancy, quality) are likely to hold true even as absolute risk levels change over time.
This study shows that, although rare, nursing home closures can be predicted with high accuracy using advanced machine learning and deep learning techniques applied to longitudinal data. Our findings highlight key risk factors that consistently suggest an increased risk of closure, including low occupancy, poor quality, and rural location. By proactively identifying at-risk facilities, predictive models can support more proactive monitoring, targeted interventions, and fair resource allocation. These tools offer a promising path to strengthening the resiliency and stability of long-term care in the United States.
